Paper recasts transformer forward pass as trajectory geometry, finding semantic convergence and 3-phase structure
its_vayishu · x · 2026-10-11
An arXiv paper proposes a probe-free mechanistic interpretability method: recasting the transformer forward pass as a discrete trajectory on a high-dimensional representation manifold, measured via trajectory length, curvature, a semantic convergence index, layerwise cosine similarity, and representational stability.
Across GPT-2, TinyLlama, and Qwen2.5, four findings:
- Semantically related prompts converge significantly in middle-to-late layers (CI 0.41–0.58, p<0.001), suggesting attractor-like dynamics
- Reasoning tasks yield higher trajectory curvature (0.71–0.83 rad vs 0.27–0.31 rad), hinting curvature encodes computational complexity
- Ambiguous tokens show trajectory bifurcation, with representational separation up to 5.6x by the final layer
- Layerwise cosine similarity reveals a universal three-phase structure: encoding, elaboration, and output preparation
Related event: New Paper Models Transformer Layer Representations as Trajectory Geometry(2 posts)→
More from Research
- EMNLP paper: diffusion LMs internally encode latent denoising timestep representations — gsarti_ · 2026-10-23
- KU Leuven and VIB open two Group Leader professorships in AI for Biology, start 2027 — jpillowtime · 2026-10-23
- Subspace Tuning curbs LLM catastrophic forgetting by freezing principal singular directions — scottniekum · 2026-10-23
- Perceptic launches PercepticOS, an AI workspace for 'scientific superintelligence' — nathanbenaich · 2026-10-23
- EPFL's Bosselut to talk long context as test-time learning at ELLIS seminar — andre_t_martins · 2026-10-23
- DRATS: Adaptive Task Sampling Boosts Worst-Task Performance in Multi-Task RL — Cohere · 2026-10-23